Caliby

Structure-conditioned sequence design—also called inverse folding—asks for an amino-acid sequence that will fold into a given target backbone, and it underpins much of modern de novo protein design. State-of-the-art models such as ProteinMPNN nonetheless struggle with many non-idealized backbones under standard in-silico success criteria. Caliby, from the Protein Design Lab at Stanford University and released as a preprint in 2025, argues that a common training objective is partly to blame: optimizing for native sequence recovery pushes models to reproduce non-structural signals such as phylogenetic relatedness, neutral drift, and dataset sampling biases rather than a clean, generalizable structure-to-sequence mapping.

Caliby is a Potts-model-based sequence design method that conditions on an ensemble of structures rather than a single backbone. It generates a synthetic conformational ensemble from an input backbone using Protpardelle-1c partial diffusion, then samples sequences consistent with the shared structural constraints of the ensemble. Averaging over the ensemble suppresses biases toward any one native sequence while preserving the structural signal common to all members.

The result is a design method that trades raw native sequence recovery for markedly better structural self-consistency, extending the reach of inverse folding to less idealized backbones.

#Key Features

• Ensemble conditioning: Designs sequences against a synthetic structural ensemble rather than a single static backbone, averaging out non-structural biases.
• Potts-model formulation: Learns a probabilistic sequence model from which designs are sampled, with dedicated modules for scoring and sidechain packing.
• Improved self-consistency: Reduces native sequence recovery while substantially improving AlphaFold2 self-consistency, outperforming ProteinMPNN and ChromaDesign on both native and de novo backbones.
• SolubleCaliby variant: A version trained only on soluble proteins recovers Protpardelle-1c binder designs that SolubleMPNN had deemed undesignable, exposing limitations of existing filtering pipelines.
• Open weights and tooling: Released under Apache-2.0 with checkpoints on Hugging Face, plus a command-line interface, a Python API, and a Colab notebook.

#Technical Details

Caliby models protein sequences with a Potts (Markov random field) formulation and conditions the learned couplings on an input structural ensemble; sequences are then drawn by sampling from this model, with separate sidechain-packing modules for downstream evaluation. The released checkpoints include the default caliby (trained on all PDB chains with 0.3 Å noise), soluble_caliby (excluding transmembrane proteins), and caliby_distill (a distilled variant that avoids the ensemble-generation step), alongside sidechain packers trained at several noise levels. Weights download automatically on first run. In-silico, ensemble-conditioned design improves AlphaFold2 self-consistency over ProteinMPNN and ChromaDesign despite lower native sequence recovery.

#Applications

Caliby is intended for protein engineers and de novo design teams who need sequences for challenging, non-idealized backbones—cases where recovery-optimized inverse-folding models tend to fail self-consistency filters. Its ensemble-conditioned sampling is particularly useful for binder design and other tasks where backbones deviate from highly regular topologies, and the SolubleCaliby variant can rescue designs that soluble-specific filters would otherwise reject. The Apache-2.0 release, CLI, Python API, and Colab make it straightforward to slot into existing design pipelines.

#Impact

By reframing inverse folding around structural ensembles and away from native sequence recovery, Caliby offers evidence that recovery-centric objectives encode undesirable biases, and it expands the designable space beyond highly idealized backbones. As a successor to ProteinMPNN and ChromaDesign with fully open weights and code, it provides a practical new tool and a conceptual shift for structure-conditioned design. As a preprint awaiting peer review, its advantages are so far demonstrated through in-silico self-consistency benchmarks rather than experimental characterization.